Fix issues causing test flakes.

* Make the tf_job_simple test robust to test flakes due to problems initializing the ksonnet app.

* Skip install the default registries because we don't need them and talking
  to Git just creates a source of flakiness.

* Add retries to setting up the ksonnet app.

Fix kubeflow#1128

* Fix flake in minikube test

wait_for_operation should just rety on socket errors.

Fix kubeflow#1137

testing/tf_job_simple_test.py

@@ -23,7 +23,8 @@
 import argparse
 import logging
 import os
-
+import re
+import subprocess
 from kubeflow.testing import test_helper, util
 from retrying import retry
 
@@ -67,12 +68,33 @@ def wait_for_tf_job():
     raise Exception("Could not find services with label tf_job_name=mycnnjob")
   logging.info("Found services with label tf_job_name=mycnnjob")
 
+@retry(stop_max_attempt_number=3)
 def test_tf_job_simple(test_case): # pylint: disable=redefined-outer-name
   args = parse_args()
-  util.run(["ks", "init", "tf-job-simple-app"])
+  try:
+    util.run(["ks", "init", "tf-job-simple-app", "--skip-default-registries"])
+  except subprocess.CalledProcessError as e:
+    # Keep going if the app already exists. This is a sign the a previous
+    # attempt failed and we are retrying.
+    if not re.search(".*already exists.*", e.output):
+      raise
+
   os.chdir("tf-job-simple-app")
-  util.run(["ks", "registry", "add", "kubeflow", args.src_dir + "/kubeflow"])
-  util.run(["ks", "pkg", "install", "kubeflow/examples"])
+  try:
+    util.run(["ks", "registry", "add", "kubeflow", args.src_dir + "/kubeflow"])
+  except subprocess.CalledProcessError as e:
+    # Keep going if the registry has already been added.
+    # This is a sign the a previous attempt failed and we are retrying.
+    if not re.search(".*already exists.*", e.output):
+      raise
+
+  try:
+    util.run(["ks", "pkg", "install", "kubeflow/examples"])
+  except subprocess.CalledProcessError as e:
+    # Keep going if the package has already been added.
+    # This is a sign the a previous attempt failed and we are retrying.
+    if not re.search(".*already exists.*", e.output):
+      raise
 
   if args.tf_job_version == "v1alpha2":
     prototype_name = "tf-job-simple"

testing/vm_util.py

@@ -3,6 +3,7 @@
 import datetime
 import logging
 import os
+import sock
 import ssl
 import subprocess
 import time
@@ -45,6 +46,9 @@ def wait_for_operation(client,
       else:
         op = client.globalOperations().get(project=project,
                                            operation=op_id).execute()
+    except socket.Error as e:
+      logging.error("Ignoring error %s", e)
+      continue
     except ssl.SSLError as e:
       logging.error("Ignoring error %s", e)
       continue
@@ -77,28 +81,28 @@ def wait_for_vm(project, zone, vm, timeout=datetime.timedelta(minutes=5),
       return
     except subprocess.CalledProcessError:
       pass
-    
+
     if datetime.datetime.now() > endtime:
       raise util.TimeoutError(
         ("Timed out waiting for VM to {0} be sshable. Check firewall rules "
          "aren't blocking ssh.").format(vm))
 
     time.sleep(polling_interval.total_seconds())
-     
+
 def execute(project, zone, vm, commands):
   """Execute the supplied commands on the VM."""
   util.run(["gcloud", "compute", "--project=" + project, "ssh",
             "--zone=" + zone, vm, "--", " && ".join(commands)])
-           
+
 def execute_script(project, zone, vm, script):
   """Execute the specified script on the VM."""
 
   target_path = os.path.join("/tmp", os.path.basename(script) + "." + uuid.uuid4().hex[0:4])
-  
+
   target = "{0}:{1}".format(vm, target_path)
   logging.info("Copying %s to %s", script, target)
   util.run(["gcloud", "compute", "--project=" + project, "scp",
             script, target, "--zone=" + zone])
-  
+
   util.run(["gcloud", "compute", "--project=" + project, "ssh",
             "--zone=" + zone, vm, "--", "chmod a+rx " + target_path + " && " + target_path])